Prototype Development Update: Initial Model Exhibits False Positives

Overview This week, we successfully developed the first working prototype of our deepfake detection system. The application integrates a user-friendly frontend with a functional backend to accept images and predict their authenticity.

The prototype uses the following tech stack:

Frontend: HTML/CSS with Bootstrap for responsive design
Backend: Flask (Python) for server-side logic
Model: Neural network built using Keras
Dataset: A combination of real and deepfake images from publicly available datasets

Prototype Features

Upload interface for image input
Preprocessing using Error Level Analysis (ELA)
Inference via a trained Keras-based neural network model
Result display with classification label (Real or Fake) and confidence score

Observed Issue: False Positives During testing, we observed that the prototype incorrectly classified some genuine images as fake, resulting in false positives. This misclassification indicates that the model might be:

Overfitting to certain compression patterns present in the training dataset
Not generalizing well to unseen image types or resolutions
Sensitive to variations in lighting, camera type, or file metadata

Planned Remediation Steps To address this issue, the following actions are planned:

Dataset Augmentation: Expand the dataset with more diverse and high-resolution real images
Model Tuning: Apply regularization techniques, and re-evaluate the model architecture
Threshold Adjustment: Fine-tune the confidence threshold for classification